I should declare my interest before anything else. I build AI systems that are not large language models, so public money going toward architectural diversity would benefit me. Everything below is written so you can check it yourself against the published document rather than take my word for it.
Australia's National AI Plan is a serious piece of work in places. The adoption diagnosis is real and measured: 29% of regional organisations have adopted AI against 40% in metropolitan areas, around 40% of First Nations people and one in five Australians remain digitally excluded. Ruling out a text and data mining exception in copyright law is a decision with teeth. Indigenous data sovereignty is treated substantively rather than as a box-tick. Technology-neutral regulation built on existing law, with a new AI Safety Institute watching, is a more mature position than the risk-tiering other jurisdictions have committed to.
And there is a hole in the middle of it.
Thirty-seven pages, no architecture
I ran a full-text terminology audit over the published PDF. Across roughly 11,900 words, here is what does not appear:
Architecture. Transformer. Large language model. Foundation model. Small language model. Mixture of experts. Neural network. Deep learning. Parameter. Inference. Scaling law. Quantisation. Distillation. Fine-tuning. Open source. Open weights. On-device. GPU.
Every one of those returns zero. "Machine learning" appears once, describing Microsoft Custom Vision analysing drone imagery in Kakadu. "Chips" appears once. "Edge" appears once, inside the phrase "cutting-edge technology". The word "model" turns up twelve times and is always an undifferentiated mass noun: AI models, locally developed models, bespoke models.
The Plan adopts the OECD definition of an AI system, which is deliberately architecture-neutral. That is the right call for setting regulatory scope. It is not sufficient for an industrial strategy, because that definition cannot distinguish a model that costs hundreds of millions of dollars to train from one that runs on a processor you could lose in a pocket. Every section downstream inherits the flatness.
Here is what that costs. A document with no concept of architecture cannot ask whether the compute demand it plans around is a fact about AI, or a fact about one way of doing AI. It cannot pose the question. So it doesn't.
Who told them this was a good idea
The Plan rests on ten references. Not one is peer-reviewed.
Four are the department citing itself or work it commissioned: two National AI Centre publications, Jobs and Skills Australia, and the Fifth Quadrant Responsible AI Index, which was sponsored by the National AI Centre. Two are genuinely independent institutional sources, AEMO and the OECD. One, the Australian Digital Inclusion Index, is real university-led research from RMIT and Swinburne, corporate-funded by Telstra.
The remaining three carry the Plan's biggest claims, and each comes from a party with money riding on the answer.
The claim that Australia ranked second globally as a data centre investment destination, which anchors two of the Plan's nine actions, is sourced to Knight Frank. Knight Frank runs a dedicated data centre practice providing brokerage and consultancy, acting for tenants, operators, landowners and investors, and reports trading around US$2 billion of assets in a twelve-month period. The Commonwealth is citing a transaction intermediary as evidence for a policy that generates transactions.
The claim that AI triage could lift stage 1 lung cancer detection from 16% to 64% is sourced to the Tech Council of Australia, an industry association founded in 2021 whose members include AWS, Google, Microsoft and Atlassian. That is a clinical claim about a national screening programme, and it is not sourced to medical literature.
And the Plan's opening credential, its first substantive bullet, establishes Australia's AI standing by reporting that we rank third globally in the use of Claude, citing Anthropic's own economic index. A sovereign capability document opens by using a foreign vendor's telemetry about its own product as a measure of national achievement. On that metric, the more dependent we become, the better we are doing.
None of this means the numbers are wrong. Knight Frank's property data is used across the sector. The point is narrower and worse: a national industrial strategy has been assembled without a single peer-reviewed citation, and where it makes commercially consequential claims it leaned on people who profit from those claims being believed. Action 8 of the same document asks Australian business to adopt "transparent, fair, and accountable" practice with "clear documentation". The standard the Plan applies to its own evidence is below the one it asks of industry.
The compute promise, and what is behind it
The Plan states the problem correctly:
"Ensuring that local companies and researchers have access to the high-performance computing power they need for AI development is critical to building a strong, local technology sector."
Here is the entire corresponding commitment:
"Mapping compute: The government is undertaking work to assess the landscape of available compute infrastructure to identify gaps and guide future investment opportunities."
Critical problem, answered with a survey. No allocation, no capacity, no guaranteed research access.
What makes this worse is what the document never mentions. Search the full text for NCRIS, National Computational Infrastructure, Pawsey, Setonix, Gadi, or supercomputer and you get zero, every time. CSIRO appears once, as a partner in the Kakadu weed-control case study. Universities appear once, in a list of workplaces that need AI skills.
Australia already runs national computational infrastructure. The Commonwealth funds it. A National AI Plan proposing to go and find out what compute we have, without naming the compute we own, is not a compute policy. It is an admission that one has not been written yet.
The substitute on offer is the more than $100 billion in announced data centre investment. That substitution does not survive inspection. Hyperscale capacity is commercial cloud sold at market rates to whoever pays. Putting it in Sydney does not give an Australian researcher compute. It moves the meter closer. The Plan treats compute located in Australia and compute available to Australians as the same thing from beginning to end, and the gap between those two is the whole sovereignty question.
Nobody costed either side
The Plan's complete treatment of building our own models is one paragraph, hedged twice: the government "is also considering the role of Australian-developed models", which "may provide benefits". No scope, no architecture, no dollars.
So let me supply the numbers it left out.
Building at the frontier is already impossible for us, and that is the useful finding rather than a discouraging one. The amortised cost of training the most compute-intensive models has grown at roughly 2.4x per year since 2016 [1]. Nothing Australia can plausibly appropriate catches that curve. Which means a plan whose only mental image of "a model" is a frontier-scale one has quietly written itself out of model development altogether.
But a sovereign model is not a frontier model. Singapore's National Multimodal LLM Programme was announced at S$70 million, roughly US$52 million, covering model development, multimodal extension and researcher access to high-performance computing. Whatever you think of that programme, it establishes the order of magnitude: a nationally directed model effort is a tens-of-millions proposition. Australia's $466 million in existing AI funding is more than enough to attempt one. It has simply not been pointed at that.
And nobody priced the rent. There is no figure anywhere in the Plan for what Australian institutions currently pay foreign providers for cloud and model services, no projection of that cost as adoption scales, and no comparison against building domestic capability. That is the central economic question of an adoption-led strategy, and it is not asked once. I have not filled the gap with an estimate, because commercial rates vary by contract and I could not verify a representative figure at source. The missing calculation is the finding: we have committed to a dependency without pricing it.
Efficiency, as an air conditioning problem
The Plan's own number: data centres drew about 4 TWh on the National Electricity Market in 2024, roughly 2% of grid-supplied power, and AEMO expects that to triple by 2030. That is the only forward-looking figure with a date attached anywhere in the document. The single quantified commitment in Australia's National AI Plan is a forecast of how much more electricity it will consume.
Globally the direction is the same. Data centres used around 415 TWh in 2024, about 1.5% of world electricity, heading for roughly 945 TWh by 2030 and 1,200 TWh by 2035, having grown at about 12% a year since 2017, more than four times the rate of electricity demand overall [2]. Peer-reviewed work reaches the same conclusion [3]. Water tracks it: training a single large model can directly evaporate hundreds of thousands of litres of freshwater [4], and industry analysis puts direct cooling water at 222 billion litres in 2025 heading toward 644 billion by 2030 [5].
Now look at the Plan's response. Liquid cooling. Closed-loop systems. Near-zero water consumption. Renewable procurement. National data centre principles covering sustainability and cooling.
Every one of those is worth doing, and not one of them reduces the amount of computation performed. The Plan optimises the building around the computation with real care, and never asks whether the computation has to be that size. Efficiency, in this document, is a cooling specification.
That would be defensible if the demand curve were a law of nature. It isn't. It is the aggregate of architectural choices, and the peer-reviewed literature is unambiguous that those choices move the number a long way.
The second option the Plan cannot name
The Plan says Australia's competitive edge is in healthcare, agriculture, resources and advanced manufacturing. Look at what those actually are: sensing, control and real-time monitoring, frequently in places with constrained power and unreliable connectivity. That is precisely the class of problem dense transformers suit worst and cost most.
Here is what is already established in the literature, none of it speculative, none of it requiring Australia to invent anything:
- Compute-optimal training. A 70-billion parameter model trained on four times more data beats models several times larger at the same compute budget [6]. Capability per dollar is a design decision, not a budget line.
- Sparse mixture of experts. Scale the parameter count while activating only a fraction per token, cutting compute and communication cost [7].
- State-space models. Five times the inference throughput of transformers, linear rather than quadratic scaling in sequence length, and inference memory that stays constant no matter how long the stream runs. A 3B model matches transformers twice its size [8]. That is a description of a sensor-stream engine.
- Neuromorphic processors. Orders-of-magnitude improvements in energy-delay product on suitable workloads [9][10]. Built for exactly the remote, power-constrained deployment Australia has more of than almost anyone.
- Event-based vision. Microsecond latency, enormous dynamic range and low power, by reporting brightness change instead of frames [11]. Mine sites, harvesters, drones.
- Computational mechanics. Causal-state reconstruction gives you the minimal model that predicts a process as well as it can be predicted [12], and its internal states are discrete and inspectable, which matters enormously under any regulator that wants to know why a decision was made.
I am not arguing Australia should abandon large models. They are the right tool for language, documents, code and images, and the Plan's adoption programmes are well aimed at exactly that. I am arguing that a national strategy which cannot name a second option has foreclosed one, by omission rather than by analysis.
So what is it, actually
There are no targets in this document. Zero occurrences of baseline, KPI, metric, milestone or benchmark. All nine appearances of "target" are the adjective "targeted". The measurement section nominates three data sources, hedged with "may include, but are not limited to", and every one of them measures adoption. None measures capability.
A plan with no baseline cannot be assessed, and a government cannot be held to it. It also makes the gap self-sealing: if the only thing you measure is how fast Australians adopt AI built elsewhere, then adopting AI built elsewhere is the only thing the system can see as progress.
Put it together and the strategy is legible, even though it is never stated. Australia will be a good host (land, power, submarine cables, fast approvals) and an enthusiastic consumer (adoption programmes, skills, third in the world for Claude). Public capability funding is $466 million, and the Plan says plainly it is "existing funding already available or committed". Announced private data centre capital is north of $100 billion. In a document claiming 1,500 Australian AI companies, exactly three are named, and two of them are data centre businesses.
That is a landlord strategy. It might even be the right one for a country our size. But the Plan never says it, never tests it, never costs it against the alternative, and simultaneously dresses it in builder language: sovereign capability, Future Made in Australia, "a leader in AI innovation and a trusted exporter of AI computing power, not just a consumer of AI technologies built elsewhere".
Read that last phrase again. An exporter of computing power. Not of models. The document is internally consistent. It is describing a rental business and calling it sovereignty.
An undeclared strategy is worse than a wrong one, because nobody can argue with it. And the thing I would most like to argue about is the one the Plan has no words for: whether a country with abundant renewables, constrained capital and world-class research institutions might do better building differently than renting harder.
References
- Cottier B, Rahman R, Fattorini L, Maslej N and Owen D (2024). The rising costs of training frontier AI models. arXiv:2405.21015. Preprint, not peer-reviewed; cited as indicative because no peer-reviewed equivalent was identified.
- International Energy Agency (2025). Energy and AI. IEA, Paris.
- de Vries A (2023). The growing energy footprint of artificial intelligence. Joule, 7(10), 2191-2194.
- Li P, Yang J, Islam MA and Ren S (2025). Making AI Less 'Thirsty'. Communications of the ACM, 68(7).
- Rystad Energy (2026). Data center water consumption could triple by 2030. Industry research.
- Hoffmann J, Borgeaud S, Mensch A, et al. (2022). Training Compute-Optimal Large Language Models. Advances in Neural Information Processing Systems 35 (NeurIPS).
- Fedus W, Zoph B and Shazeer N (2022). Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. Journal of Machine Learning Research, 23.
- Gu A and Dao T (2024). Mamba: Linear-Time Sequence Modeling with Selective State Spaces. First Conference on Language Modeling (COLM).
- Davies M, Srinivasa N, Lin T-H, et al. (2018). Loihi: A Neuromorphic Manycore Processor with On-Chip Learning. IEEE Micro, 38(1), 82-99.
- Davies M, Wild A, Orchard G, et al. (2021). Advancing Neuromorphic Computing With Loihi: A Survey of Results and Outlook. Proceedings of the IEEE, 109(5), 911-934.
- Gallego G, Delbruck T, Orchard G, et al. (2022). Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1), 154-180.
- Shalizi CR and Crutchfield JP (2001). Computational Mechanics: Pattern and Prediction, Structure and Simplicity. Journal of Statistical Physics, 104, 817-879.
- Department of Industry, Science and Resources (2025). National AI Plan. Australian Government. All quotations and term counts in this article are taken from this document.